One of the more notable details from the shipping manifests is the existence of BGA 1074, which implies that the product being mentioned here is indeed an APU, rather than another configuration. Another interesting mention is the package size, which measures 32mm x 27mm, placing it in the mobile SoC territory and optimizing it for OEM integration. Another reason the manifests here indicate an entirely new lineup is the 0.8 mm pitch and the mention of AMD’s new FF5 socket, which replaces the FF3 socket, commonly found in Valve’s Steam Deck.
Rumors about AMD’s Soundwave APUs emerged a few months ago, around the same time NVIDIA was reported to be developing an AI PC chip based on the ARM architecture….
NVDA’s current Grace CPU is a 72 high performance ARM neoverse V2 cores and uses LPDDR5X memory with ECC supporting up to 480 GB with %!@ GB/s bandwidty. The Blackwell GPU is built on TSMC 4nm process with multi chip module design combining 2 large dies and has 192 GB of HBM3e memory (who makes this HBM). Got this off of AI…doc
The answer I’m finding is yes it’s an AI oriented PC/data center chip based on the Arm architecture. It supports tight coupling with GPUs for accelerated computing and uses advanced memory subsystems like LPDDR5X. AMD apparently does not have a cpu that directly competes with these ARM cpu’s interestingly…doc
I’m not sure what Soundwave is and I’m not sure the Grace ARM is for home computers and data centers like AMD cpu’s. From my reading I would say no because the Grace cpu is made for primarily the Nvidia gpu and AI uses…doc
Thanks. So is that not already an “AI PC chip based on the ARM architecture”? Or is it not a PC chip?
Neoverse is purely a server CPU design. Fairly widely used rather than players designing their own Arm cores for server – for example, the in-house Arm chips at Amazon, Google, Microsoft Azure and some others are all Neoverse derivatives. Nothing to do with personal computing or mobile.
I hadn’t realized that Nvidia was using Neoverse designs in their CPU/GPU combos for servers like Grace Hopper. If they are, they’re less differentiated in CPU than I thought, though obviously integrating a competent server CPU with Nvidia’s best GPU for servers/AI would be incredibly powerful..
This is a good interview and very informative. It makes me want to get some NVDA because moving forward, they seem to control their destiny by owning most of the aspects of AI with NVDA products from hardware to software. I’m a big AMD fan but It seems that NVDA still has a moat while AMD is doing great as well, but going with more open source software gives them an advantage in cost but also guarantees them a possibly smaller profit on their systems. I think they are both good investments with NVDA staying number one and AMD continuing to erode market share from NVDA. IMHO…doc
It’s been weird to watch the Arm server CPU players basically all flame out. Very weird. The one that has really come through with something great is the Nuvia Oryon which has of course transformed into the recent Qualcomm Snapdragon generations for mobile and was abandoned for servers. So all the money and effort that went into those designs basically went nowhere. And we have Neoverse which obviously must be pretty decent if it’s being so widely adopted.
If literally everyone who took the Architecture License Agreement for servers failed… were Arm licenses for servers bordering on a scam of some kind? (Joking, but only half joking. Dollars flow in, nothing flows out?) Or was the hill just THAT steep to climb in terms of software ecosystem, growth of the cloud players choosing their own CPUs instead of third party, etc.?
Will AMD even try to produce an ARM server CPU eventually? They have the means, where no one else does, since they can start on top of the existing Zen designs. But maybe there’s no point. I’m not sure who would ever adopt it, unless Oracle gets impatient with Ampere and still has reason for sticking with a non-Neoverse ARM. And Soundwave is certainly not this.
Will Soundwave ultimately be based on something like a Zen 5c back end plus ARM instruction decoding? Or will it just be a licensed ARM core plus AMD GPU tech, shrunk down? I’m betting on the former, but who knows?
I remember that it was a heavy lift to get, for example, software to take advantage of Optane, this was one of the reasons Optane and that entire accompanying CPU generation didn’t go very far in the data center. But if you’re just moving from one generation of a CPU core to another, without major instruction set changes, maybe it isn’t that hard. I’m trying to assess whether all the Arm software ecosystem work that’s happened on the Neoverses would just carry over to another instruction-set-compatible CPU from another vendor. Must be some level of ready compatibility plus or minus performance tuning, or Oracle would have little use for Ampere.